Why Your AI Bill Keeps Growing While Token Prices Keep Falling

The Founders’ AItrained on 25 years of our work

Want this made concrete for your company? Ask me — or pick one:

Why is our AI bill exploding — and how do we get it under control? What should I do first? Take the 90-second readiness read

Ask

In June 2026, TechCrunch reported that Uber had burned through its entire annual AI tooling budget in roughly four months. The company’s response was a hard cap: $1,500 per employee, per month, per agentic coding tool.

Here is what makes that story worth your attention: it happened while AI got cheaper. Every major lab cut or held per-token prices over the past year — Anthropic’s newest Sonnet tier launched below its predecessor’s price, OpenAI’s budget tiers run under a dollar per million input tokens, and all three major providers now publish 50% batch discounts and 90% caching discounts. The unit price of intelligence is falling. Total bills are rising anyway.

That contradiction is the real story of AI cost in 2026, and most budget conversations miss it. The companies overrunning their AI budgets — 79% of enterprises in the past twelve months, according to a February 2026 survey of 500 finance leaders — are not victims of expensive tokens. They are running exploding volume through an architecture nobody designed. Usage shifted from chat to agents, agents consume tokens at a fundamentally different rate, and finance found out when the invoice arrived.

The fix is not panic, and it is not turning the AI off. The productivity is real — Uber kept spending, deliberately, after the overrun. The fix is a set of specific, unglamorous architecture and governance decisions that most companies simply haven’t made yet. This guide covers what AI actually costs in 2026 — tokens and seats, verified against every provider’s published pricing — why bills explode, and the five levers that control them. And because the physics are the same but the failure modes are different, we’ll run the math at two scales throughout: a 25-person company and a 1,000-person company.

The Bill Went Up While the Price Went Down

Start with the number that explains all the others. In April 2026, a research team including Erik Brynjolfsson and Alex Pentland published a study with a title that sounds like a joke and reads like a warning: “How Do AI Agents Spend Your Money?” Their finding: agentic tasks — where the AI works in multi-step loops, reading files, calling tools, checking its own output — consume roughly 1,000 times more tokens than asking the same model a question in chat.

Not 30% more. Not double. Three orders of magnitude.

Two details from that study matter even more for budgeting than the headline multiplier. First, the cost is driven by input tokens, not output — agents re-read their accumulated context on every step, so the longer a task runs, the more each subsequent step costs. Second, runs of the identical task varied by up to 30x in token consumption. The same request, the same model, a 30x cost spread. You cannot forecast that with a spreadsheet built for per-seat software — and the study found that higher token burn didn’t even buy better results: accuracy peaked at intermediate cost and then flattened.

This is not a niche coding phenomenon. The a16z and OpenRouter “State of AI” report — built on more than 100 trillion tokens of real usage across 300+ models — names agentic inference the fastest-growing usage pattern in the market, with OpenRouter alone processing over a trillion tokens a day. The whole market is shifting from the cheap shape of AI usage to the expensive one.

Now the Uber story makes sense. Per Forbes’ reporting, adoption of agentic coding tools inside Uber went from 32% of engineers in February to 84% in March — one month. Roughly 70% of committed code originated from those tools. Average cost ran $150 to $250 per engineer per month, but power users ran $500 to $2,000 — and Uber’s own CTO, Praveen Neppalli Naga, reportedly spent $1,200 in a single two-hour demo session. Multiply a fast-moving adoption curve by a 1,000x workload multiplier and a 30x variance band, and a well-planned annual budget lasts four months. Not because anyone did anything wrong — because the consumption model changed underneath the budget model.

One more piece of context, because precision matters here: enterprise AI spend is growing because companies choose to grow it. Menlo Ventures measured $37 billion in enterprise generative AI spend in 2025, up 3.2x from $11.5 billion the year before. And Bain’s June 2026 survey of 951 companies found that 90% of companies whose AI cost savings underperformed are increasing their AI budgets anyway. The market has concluded the value is real. What it hasn’t built yet is the cost architecture underneath that conviction — which is exactly the gap this guide is about.

What AI Actually Costs in 2026: Tokens and Seats

AI spend arrives on two meters, and they behave completely differently. Seats are predictable and per-person: chat assistants, copilots, workspace AI. Tokens are usage-based and per-task: API calls, agents, automated workflows. Most companies now pay both, and most budget surprises come from treating the second meter like the first.

The token prices below are from each provider’s official pricing page, verified July 2026, per million tokens:

Read the table vertically and one pattern jumps out: within each provider’s current family, the frontier tier costs roughly five to ten times more per token than the small tier. That ratio is the single most useful number in this guide, because most day-to-day business tasks — summarization, classification, drafting, extraction, support triage — do not need frontier reasoning. Every routine task running on a frontier model is paying a 5–10x premium for judgment it isn’t using.

Anthropic’s own documentation makes the point with a worked example: processing 10,000 customer support conversations on its small model costs about $37 total. The same workload routed to its frontier tiers would run five to ten times that — $185 to $370 by the same arithmetic. Neither number breaks a budget. But scale that decision across every workflow in a company, every day, and the model-tier decision quietly becomes the largest single lever in your AI cost structure.

On top of the base rates, every major provider publishes two standing discounts that most companies leave unclaimed. Prompt caching cuts the cost of repeated context — system prompts, reference documents, conversation history — by 90%: a cache read costs one-tenth of the standard input price at Anthropic, OpenAI, and Google alike. Since agentic workloads are input-heavy by nature (that’s the arXiv finding), caching attacks exactly the tokens that dominate agent bills. Batch processing gives a flat 50% discount on anything that doesn’t need an immediate response — overnight reports, bulk classification, content pipelines — again at all three providers.

The seat side is simpler, and cheaper than most executives assume:

(OpenAI does not publish a public list price for its business tiers; reported figures put ChatGPT Business in the same $20–25 band, with enterprise agreements individually negotiated.)

Notice what Anthropic did with its enterprise plan, because it’s a signal about where the whole market is heading: the seat is $20, and usage bills separately at API rates. The vendor itself has concluded that flat per-seat pricing can’t represent agentic usage — one seat’s consumption is no longer predictive of another’s. Keep that structure in mind; it’s the pricing model catching up to the 30x variance finding, and it previews the seats-versus-tokens decision later in this guide.

The 25-Person Bill and the 1,000-Person Bill

Same physics, different failure modes. Run the arithmetic at both scales and you can see exactly where each one’s budget breaks.

At 25 people, the seat baseline is small. Twenty-five seats at the published list prices above lands between roughly $350 per month (Google Workspace Standard) and about $550 per month (Claude Team, billed monthly) — call it $4,000 to $7,000 a year. That is not a number that needs a committee.

What breaks a 25-person budget is not the seats. It is the moment one or two people discover agentic workflows. Recall Uber’s disclosed range: power users of agentic coding tools ran $500 to $2,000 per month each. At a 25-person company, a single enthusiastic power user can quietly cost one to four times as much as the entire company’s seat bill — and there is usually no meter anywhere that makes this visible until the card statement arrives. The 25-person failure mode is concentration: a handful of ungoverned power workflows sitting invisibly on top of a small, predictable base.

At 1,000 people, both meters are large and both misbehave. The seat math alone deserves attention: 1,000 seats at $20 is $20,000 a month — $240,000 a year — before anyone runs a single agentic task. At that line size, seat utilization becomes a real budget question: seats bought ahead of adoption sit idle while the invoice runs, which is why a utilization audit belongs in the quarterly rhythm of any large deployment (we cover the adoption side of that problem in why your team isn’t using AI).

Then the usage meter arrives on top. Suppose agentic tools reach just 10% of a 1,000-person company — one hundred people — at Uber’s average of $150–$250 per user per month. That is another $15,000 to $25,000 a month, roughly doubling the AI line, from one-tenth of the workforce. Now recall how fast that percentage moves: Uber went from 32% to 84% adoption in a single month. The 1,000-person failure mode is compounding: seat sprawl, plus shadow usage on personal accounts that never touches the meter at all — a risk we’ve mapped in the shadow AI problem — plus agentic multiplication spreading team by team, each dynamic amplifying the others.

The lesson from both columns is the same and worth stating plainly: AI cost is not a function of headcount. It is a function of workload shape. A 25-person firm with three heavy agentic workflows can out-spend a 200-person firm using chat assistants. Which is why the control levers that follow are about workloads, not seats.

Five Cost Drivers, Five Levers

Nearly every exploding AI bill we see traces to some combination of five drivers. Each one has a specific, published, verifiable lever — none of them require abandoning a workload, and none of them require negotiating with a vendor.

1. Everything runs on the top model → orchestrate instead

The most common driver is also the simplest: everything runs on the frontier model because that’s what someone set as the default. Using the top model for everything feels like a quality decision. It’s actually an unexamined one — the frontier-to-small price gap is 5–10x per token at every major provider, and the research on matching models to tasks is unambiguous. RouteLLM — the foundational routing study from Berkeley’s LMSYS group — showed a learned router achieving 95% of GPT-4’s quality while sending only 26% of calls to GPT-4, and with better training data, only 14% — an overall cost reduction of more than 85% on its main benchmark.

The mature version of this lever is model orchestration: instead of one model doing everything, workflows are designed so each tier does what it’s priced for. A frontier model handles the judgment steps — planning the task, making the calls that require reasoning, reviewing the final output — while small models handle the volume steps: extraction, classification, summarization, formatting, the repetitive middle of every agentic loop. The vendors themselves design for this — Anthropic’s own pricing documentation recommends exactly the split: the small tier for simple tasks, the mid tier for most production workloads, the frontier tier for the hardest reasoning. Since agentic tasks burn tokens in the repetitive middle (that’s where the 1,000x goes), orchestration puts the 5–10x price gap to work on precisely the tokens that dominate the bill — without lowering the quality ceiling, because the frontier model still owns every step that needs it.

The practical version doesn’t need a research-grade router. At 25 people it’s a one-page policy: small model by default, frontier by exception, and any agentic workflow that runs on the top model end-to-end gets redesigned. At 1,000 it’s configuration in your AI platform — tier defaults per task class, owned by whoever owns the platform, so orchestration is the path of least resistance rather than a per-team habit.

2. Context bloat → cache what repeats

Agent bills are input-token bills — the arXiv study found input, not output, drives agentic cost, because agents re-read their context at every step. The published fix is prompt caching: at Anthropic, OpenAI, and Google alike, a cached input token costs 10% of the standard rate. System prompts, policy documents, knowledge bases, long conversation histories — anything static that gets re-sent should be cached, and at Anthropic’s published multipliers the cache pays for itself after a single re-read. If your team is building on the APIs and hasn’t turned caching on, this is the highest-ratio configuration change available: 90% off the exact token category that dominates agentic spend.

3. Agentic loop multiplication → cap it, then watch it

The same task can cost 30x more on one run than another — that’s not a hypothetical, it’s the measured variance in the arXiv data. You cannot forecast that; you can only bound it. Uber’s answer was a hard cap — $1,500 per employee per month per agentic tool — and the structure of that answer matters more than the number: per employee, per tool, monthly, visible. A cap converts an unbounded distribution into a bounded one, which is what makes budgeting possible again. Set the cap generously (the point is bounding runaway loops, not rationing productivity), alert at 80%, and review the outliers monthly — the outliers are usually your most valuable power users, and what they need is a better-designed workflow, not a smaller allowance. If agents are multiplying across your organization faster than anyone designed for, that’s its own architectural problem — one we’ve written about in AI agent sprawl.

4. Everything runs real-time → batch what can wait

Half of most AI workloads don’t need an answer in seconds — nightly reports, bulk document classification, content generation pipelines, data enrichment. Every major provider prices that patience at a flat 50% discount through their batch APIs. This is the least glamorous lever on the list and the easiest to implement: an inventory of your workloads sorted by “needs an answer now” versus “needs an answer today” is, functionally, a 50%-off coupon on the second column.

5. Seats bought ahead of adoption → audit and right-tier

The seat meter fails quietly: licenses purchased for everyone, used by some. The fix is a quarterly utilization review — active users against paid seats, by tier — and right-tiering against the ladders the vendors already publish: Google Workspace runs $7 to $22 depending on how much AI access a role actually needs; Anthropic prices a standard seat at $20 and a 5x-usage premium seat at $100, which is exactly the split between an occasional user and a power user. Right-tiering is not about cutting people off — it’s about paying the power-user price only for actual power users. If you’re still choosing which platform’s seats to buy in the first place, we’ve compared the four major options in ChatGPT vs. Claude vs. Copilot vs. Gemini.

Seats or Tokens: Choosing the Pricing Model That Fits

With both meters understood, the procurement question gets simple: match the pricing model to the workload shape.

Seats fit chat-shaped work. When usage is roughly uniform per person — drafting, summarizing, asking questions, meeting notes — a flat seat price is fair to both sides and trivially predictable. A $14–$25 seat that saves an employee an hour a week doesn’t need an ROI model; it needs a signature.

Usage-based fits agent-shaped work. The moment workloads become agentic, per-person pricing stops describing reality — consumption is driven by task complexity, not headcount, and it varies 30x run to run. Usage-based billing is not a vendor inconvenience; it’s the honest meter for that shape of work. The market’s own behavior confirms the split: Anthropic’s enterprise plan now prices the seat at $20 and meters usage separately at API rates — the two shapes of work, unbundled on one invoice.

The practical guidance by scale: a 25-person company should buy seats for everyone who wants them (the baseline is a rounding error) and put its energy into governing the two or three agentic workflows that will actually dominate spend. A 1,000-person company should treat the two meters as separate budget lines with separate owners — seats managed like any SaaS line with quarterly utilization reviews, usage managed like a cloud bill with caps, alerts, and a routing policy. The mistake at both scales is the same: managing the meter you understand and ignoring the one that’s growing.

AI FinOps That Scales Down

The companies that overran weren’t the immature ones. Self-rated FinOps leaders overran 89% of the time — the discipline that’s missing is AI-specific, and it’s new for everyone.

The finance world has already named this discipline: AI FinOps. The FinOps Foundation’s 2026 survey tells you how fast it arrived — 98% of practitioners now manage AI spend, up from 63% a year earlier and 31% the year before that, and the same survey names AI cost management the number-one skill their teams need to develop. No capability in that survey’s history has moved that fast, because no cost line has moved that fast.

Here’s the finding that should reframe how you think about your own readiness: general cloud-cost maturity doesn’t protect you. In the DoiT survey, organizations that rated themselves very mature on FinOps still overran their AI budgets 89% of the time — with the highest average overspend of any group, at 30.9%. Everyone is early. The companies getting this right aren’t the ones with the biggest FinOps teams; they’re the ones that put AI-specific visibility in place before the volume arrived.

And visibility is the honest frame — not austerity. Bain’s data shows 90% of companies whose AI savings underperformed are raising budgets anyway, and Uber’s leadership treated its overrun as a signal of adoption, not waste. The discipline exists so you can keep spending with confidence — knowing which workloads produce value at what cost, instead of discovering the total on the invoice. Only 15% of finance leaders can currently calculate AI ROI without significant bottlenecks; the operating model below is what closes that gap (and if you’re working out what AI returns at each stage of maturity, the AI ROI map is the companion piece to this one).

Budgeting for what you can’t forecast

Before the checklists, face the uncomfortable part directly: a portion of AI usage cost is genuinely unforecastable, and pretending otherwise is how budgets die. The variance is structural, not a planning failure. The same agentic task can consume 30x more tokens on one run than the next — the models can’t even predict their own consumption. Adoption moves faster than any planning cycle: Uber went from 32% to 84% of engineers on agentic tools in a single month. And every new workflow your team invents resets the baseline. You cannot forecast this line the way you forecast software seats. You can only build a budget structure that absorbs what it can’t predict — four pieces:

Bound the tail. A per-employee, per-tool cap (Uber’s structure: $1,500 per month, generous, visible) converts an unbounded cost distribution into a bounded one. That single move makes everything else plannable: whatever the variance does inside the cap, the worst month has a known ceiling.

Shorten the feedback loop. Unforecastable costs are survivable when you learn about them in days and fatal when you learn at quarter-end. Alerts at 80% of any budget line, and a review cadence measured in weeks — not a quarterly finance meeting discovering a four-month-old trend.

Carry a variance reserve. Budget the usage line with an explicit buffer rather than a point estimate. If you want calibration for a first year, use the market’s own revealed error: enterprises that overran their AI budgets averaged 30.9% overspend. Reserving on that order — and treating an untouched reserve as next quarter’s experiment budget, not savings — turns the same volatility from an overrun into a plan.

Quarantine the new. Any genuinely new agentic workflow gets a sandbox budget and a 60–90 day metering period before it earns a permanent line. New workloads are where the 30x surprises live; after two months of real usage data, they become forecastable like everything else. The unpredictability never disappears — it just keeps moving to whatever you adopted most recently, which is exactly where your attention should be.

At 25 people, AI FinOps is a rhythm, not a team:

• One named cost owner — usually the ops lead or a founder. Not a committee.
• Billing alerts on every AI account, set at a monthly number you’ve actually chosen.
• A default-model policy: small model by default, frontier tier by exception.
• Fifteen minutes a month reviewing the usage dashboards the providers already give you: top spenders, top workflows, anything growing faster than adoption explains.

That’s the whole program. A spreadsheet and a calendar reminder, not a platform purchase.

At 1,000 people, the same rhythm needs structure:

• Metering from day one: usage tagged by team and workflow, so the bill decomposes into decisions.
• Showback or chargeback: teams see what their usage costs. Visibility alone changes behavior — the goal is a thousand people making informed trade-offs, not one gatekeeper making slow ones.
• A model-orchestration policy owned by whoever owns your AI platform, with the tier defaults from lever #1 baked into configuration rather than left to each team’s habits — frontier models for judgment steps, small models for volume steps, by design rather than by default.
• Per-employee caps on agentic tools, Uber-style — generous, alerted, reviewed monthly.
• An AI Enabler in each major function who owns adoption and cost for their area — because the person driving usage up should be the same person who understands what it costs.

Notice what neither list includes: a spending freeze, a tool ban, or a procurement gauntlet. The cost problem in AI is an architecture and visibility problem. Solve it that way, and the budget conversation changes from “why is this bill so high?” to “which of these workloads deserves more?”

If You’re Budgeting the Bigger Picture

Token and license spend is one line of a complete AI budget — and usually not the largest one. Two companion guides complete the picture:

If you’re budgeting an AI program end to end — consulting help included — the running costs in this guide sit alongside advisory fees, implementation, change work, and governance. Our complete AI consulting cost guide covers what every tier of firm charges in 2026 and the 20–40% of program cost that hides outside the quote; this article is, in effect, a deep-dive on one of those hidden lines.

If you’re deciding what to run in-house versus with help — the same architecture decisions that control cost (routing, caching, governance) are the ones that determine whether you end up owning your AI operating layer or renting it. The mid-market architecture engagement covers what ownership-transfer actually looks like in practice.

Sources

• How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks — Bai, Huang, Wang, Sun, Mihalcea, Brynjolfsson, Pentland, Pei; arXiv, April 2026 (agentic tasks consume 1,000x more tokens than code chat; input tokens drive cost; up to 30x variance across identical runs; accuracy peaks at intermediate cost)
• Uber caps employee AI spending after blowing through budget in four months — TechCrunch, June 2026 ($1,500/employee/month cap per agentic coding tool)
• Uber burns its 2026 AI budget in four months on Claude Code — Forbes, May 2026 (adoption 32%→84% in one month; ~70% of committed code AI-originated; $150–$250 average, $500–$2,000 power users)
• AI spending survey: 79% of enterprises report AI cost overruns — DoiT / Sapio Research, February 2026 (500 finance leaders, 1,000+ employee orgs, US/UK; 89% of self-rated FinOps leaders overran, 30.9% mean overspend; 15% can calculate AI ROI without bottlenecks)
• State of FinOps 2026 — FinOps Foundation, 2026 (98% manage AI spend, up from 63% and 31% prior years; AI cost management the #1 skill gap)
• Your AI Budget Is Growing. Your Returns Aren’t. Here’s Why. — Bain & Company, June 2026 (Automation and AI Pathfinder Survey, n=951; ~40% of measured savings below 10%; 90% increasing budgets anyway; 7% run fully autonomous agents)
• 2025: The State of Generative AI in the Enterprise — Menlo Ventures, December 2025 (495 US enterprise decision-makers; $37B enterprise genAI spend, 3.2x YoY; LLM spend share Anthropic 40%, OpenAI 27%, Google 21%)
• The State of AI — a16z / OpenRouter, December 2025 (100T+ tokens analyzed across 300+ models; >1T tokens/day; agentic inference the fastest-growing usage pattern)
• RouteLLM: Learning to Route LLMs with Preference Data — LMSYS / UC Berkeley, 2024 (95% of GPT-4 performance with 26% of calls; 85%+ cost reduction on MT-Bench)
• Claude API pricing — Anthropic, retrieved July 2026 (model rates; caching multipliers; batch 50%; 10,000-ticket worked example)
• OpenAI API pricing — OpenAI, retrieved July 2026 (model rates; 90% cached-input discount; batch 50%)
• Gemini API pricing — Google, retrieved July 2026 (model rates; context caching ~90%; batch 50%)
• Claude plans and pricing — Anthropic, retrieved July 2026 (Team $20–$25/seat; premium seats; Enterprise $20/seat + usage at API rates)
• Microsoft 365 Copilot pricing — Microsoft, retrieved July 2026 (Copilot Business add-on $18, standard $21)
• Google Workspace pricing — Google, retrieved July 2026 (Starter $7 / Standard $14 / Plus $22 per user/month)
• Gartner: Worldwide AI Spending Will Total $2.5 Trillion in 2026 — Gartner, January 2026 (market context: total AI spending $2.5T)

Frequently Asked Questions